Collaborative Threshold Watermarking

Official implementation of Collaborative Threshold Watermarking (arXiv 2602.10765).

Overview

We propose (t,K)-threshold watermarking for federated learning. K clients collaboratively embed a shared watermark during training. Only coalitions of at least t clients can reconstruct the watermark key and verify a suspect model. Secret sharing prevents verification by smaller coalitions and protects the key during verification.

The watermark remains detectable at scale (K=128) with minimal accuracy degradation and resists fine-tuning attacks with up to 20% of training data.

Setup

conda create -n fedwm python=3.11 -y
conda activate fedwm
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install -e .

Usage

Baseline (no watermark):

python example/baseline_resnet18.py          # CIFAR-10
python example/baseline_resnet18_cifar100.py # CIFAR-100
python example/baseline_tinyimagenet.py      # TinyImageNet

Trusted server:

python example/trusted_server_resnet18.py
python example/trusted_server_resnet18_cifar100.py
python example/trusted_server_tinyimagenet.py

Untrusted server:

python example/untrusted_server_resnet18.py
python example/untrusted_server_resnet18_cifar100.py
python example/untrusted_server_tinyimagenet.py

Trustless key generation:

python trustless_keygen/simulation_distributed.py
python trustless_keygen/benchmark_distributed.py

Project Structure

src/                  # Core library
  Client.py           # Base federated client
  ClientResNet18.py   # ResNet18 client (CIFAR-10/100)
  ResNet.py           # ResNet architecture
  data_utils.py       # Dataset utilities
  utils.py            # Watermark generation and verification
example/              # Experiment scripts
trustless_keygen/     # MPC-based distributed key generation

Citation

@inproceedings{bakr2026collaborative,
  author    = {Bakr, Tameem and Ambreth, Anish and Lukas, Nils},
  title     = {Collaborative Threshold Watermarking},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026},
}